JPH04148482A - Method and device for fuzzy pattern recognition - Google Patents

Abstract

PURPOSE:To execute efficient and accurate pattern recognition by comparing a graphic formed with a binary signal based on the combination of adaptation of inputted coordinate points to preliminarily set membership function with a graphic having fuzzy boundary. CONSTITUTION:The binary signal transmitted by a signal conversion part 2 is checked against graphic information by three-dimensional membership function combined in a membership function setting part 3 and the degree of their coinci dence is calculated. Such check is executed for all graphic patterns which are candidates for recognition. While treating the degree of coincidence as certainty, a signal expressing the certainty is outputted along with a signal expressing the checked shape. Thus, a fuzzy area having a desirable shaped boundary can be set within input variable space and the accurate pattern recognition can be obtained.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention utilizes fuzzy theory to evaluate figures that are not limited to exact shapes, such as handwritten characters, using figures with ambiguous boundaries. This invention relates to a method and an apparatus for carrying out the method.

[Prior Art] Conventionally known pattern recognition techniques generally target figures whose shapes are strictly defined, but this technology also accepts figures that are not limited to exact shapes, such as handwritten characters, as recognition targets. There is a method using neural networks.

In this method, a neural network is formed by learning about the recognition target in advance, and a recognition result is obtained by manually inputting the pattern to be recognized into the neural network.

However, the neural network method requires learning, and relearning is required each time the number of pattern candidates to be identified increases.The neural network is a black box. If there is a change in the characteristics of the pattern candidate to be identified due to changes over time, etc., there is a drawback that it is difficult to reflect this change in the operation of the device.

Recently, pattern recognition has been considered as one of the application fields of fuzzy theory. Conventional pattern recognition methods that apply fuzzy theory involve having knowledge about multiple features of a figure, such as ``There is a vertical line near this part,'' and checking them one by one. It is basically closed.

[Problem to be solved by the invention] However, conventional methods based on fuzzy theory require work to acquire knowledge about the characteristics of shapes (rule creation work) when developing systems including software. A feature extraction algorithm is required.

Therefore, many rules and membership functions of fuzzy sets are required, resulting in a problem that the processing becomes complicated.

Therefore, the purpose of the present invention is to achieve efficient and accurate pattern recognition using fuzzy theory, rather than using a black box such as a neural network, and without the need for learning or knowledge acquisition rules or complex feature extraction algorithms. The purpose of this invention is to provide a method and apparatus for doing so.

[Means for solving the problem] The method of the present invention performs pattern recognition based on a membership function that defines a fuzzy set. By converting the signal into a 21m signal corresponding to the image, and comparing the binary signal with a figure with an ambiguous boundary based on a combination of fitness degrees of membership functions of a fuzzy set set in advance, It is characterized by calculating a confidence level as an evaluation value of an image.

The apparatus of the present invention implements the above method, and comprises:
A signal input unit that inputs an image signal obtained by reading an image to be recognized, a signal conversion unit that converts the input image signal into a binary signal corresponding to the image, and a three-dimensional membership that defines a fuzzy set. a membership function setting unit that sets one or more functions and forms a figure with ambiguous boundaries by combining the fitness degrees of the membership functions; and a calculation unit that calculates and outputs a certainty factor as an evaluation value of the image by comparing the image and the graphic.

In the present invention, the image to be recognized may be further compared with a plurality of graphical patterns, and the patterns may be classified into patterns with the highest degree of certainty.

[Operation] According to the present invention, in a three-dimensional space in which a coordinate axis for forming a two-dimensional plane on which an image exists and a coordinate axis for giving a degree of conformity to a fuzzy set are added, a plane is formed by lines of equal conformity of an arbitrary shape. Membership functions are set in advance. Then, when an image signal corresponding to the image to be recognized is input manually, that image signal is converted into a binary signal (0°l) corresponding to the image, and a dot with a value of l is converted into a binary signal (0°l) corresponding to the image. The coordinates (coordinate points) of The figure formed by the binary signal and the figure with ambiguous boundaries are compared based on the combination of the suitability of the input coordinate points with respect to the preset membership function, and as a result, the evaluation value of the image is determined. A confidence level is calculated.

Furthermore, the input corresponding to the image is compared with other figures with ambiguous boundaries, classified into the pattern with the highest degree of certainty, and the classification result and degree of certainty are output.

In this way, m
By using the three-dimensional membership function expressed by the formula -, it is possible to set a fuzzy region with a boundary of an arbitrary shape (for example, an ellipse) in the input variable space and realize highly accurate pattern recognition.

[Example] FIG. 1 is a block diagram showing the configuration of a fuzzy pattern recognition device according to an example.

This device consists of an image signal human input section 1 that manually generates a signal representing an image to be recognized, a signal conversion section 2 that converts an input image signal into a binary signal corresponding to the image, and a signal conversion section 2 that converts an input image signal into a binary signal corresponding to the image. Set one or more three-dimensional membership functions that define a fuzzy set, and set a membership function that stores information for forming a figure with ambiguous boundaries by combining the fitness of the three-dimensional membership functions. By comparing the binary signal sent from the signal conversion section 2 and the graphical signal sent from the membership function setting section 3, the certainty factor 2 is calculated as the evaluation value of the recognition target. ,
Furthermore, there is an arithmetic processing unit 4 which classifies the pattern into the one with the highest degree of certainty compared to other figures with ambiguous boundaries and generates the recognition result, and a signal binding unit which displays the recognition result on a screen or prints it out. It also includes a recognition result output unit 5 that outputs the recognition result.

In this device, an image signal input unit 1 includes a light source that generates light for scanning an image to be recognized, and a photoelectric converter that reads the image using the light from the light source and converts the obtained optical signal into an electrical signal. It consists of:

The signal conversion unit 2 converts the image signal sent from the image signal input unit 1 into 0. ] includes an A/D converter that converts the signal into a binary signal. It also includes a correction circuit that corrects the size of the image. For example, if the image is squashed vertically, you can stretch it to ten squares.

The membership function setting section 3 sets one or more three-dimensional membership functions necessary for determining the shape to be recognized from input means 6 connected to the membership function setting section 3, and determines the membership of these functions. It consists of 1,000 memorabilia that stores how to combine functions and the extended range of fitness. In the graphic pattern formed in this way, all candidates to be recognized are stored. The contents will be explained in detail later.

The input means 6 includes an operation means such as a keyboard which is operated by the operator to adjust or change the shape to be recognized, or a display such as a CRT for display.

The arithmetic processing unit 3 is composed of a CPU programmed to execute the following operations. As will be explained in detail later, its operation is to compare the binary signal sent from the signal conversion section 2 with the graphic information based on the three-dimensional membership function combined in the membership function setting section 3, and to Calculate the degree of matching. This matching is performed for all graphic patterns that are candidates to be recognized. Then, using the degree of matching as the confidence level, a signal (
This outputs a signal representing the confidence level (shape number, shape name, etc.).

Next, we will explain the principles and methods of pattern recognition using examples 4.
Ru.

(A) Setting of Membership Function First, one or more three-dimensional membership functions are combined to set a three-dimensional membership function corresponding to the shape of the figure to be recognized.

For example, in a three-dimensional space consisting of x-y orthogonal coordinates and degrees of fitness, a membership function 'rJ having an elliptical line of equal fitness as shown in FIG. 2 is created. Here, the symbols in the figure are defined as follows.

For simplicity, the origin is assumed to be the reference point (the center point for describing the three-dimensional membership function most simply).

11x, y, Rx, Ry)・0 Function giving elliptical shape of isofitness line tx = f(x) Hanging mirror membership function in x-L plane 1゜g (y) Hanging mirror in y-jy plane Type membership function R.

Radius in the X-axis direction of an elliptical isofitness line with two degrees of fitness l Radius in the y-axis direction of an ellipse isofitness line with two degrees of fitness l: Parameter dy proportional to the fuzzy entropy of X : Parameter X proportional to fuzzy entropy for y Distance dy on the x-t plane between the isofitness line containing any point (x, y) and the isofitness line with goodness of fit l Arbitrary The distance on the -3'-jy cross section between the isofitness line including the point (x, yl) and the isofitness line with the goodness of fit l.The point (the goodness of fit of x, yl) given by the three-dimensional membership function For convenience of explanation, the fitness degree t will be described separately into 1. and 1., but 1.1x and 1y are practically the same coordinate axes.

In this case, the equal compatibility line including the arbitrary points fx, y) is as follows.

Also, 1.

The membership function for 1 is It is expressed as

here, - for decoding is X.

y is positive The part of + is X y represents the negative side part.

Focus on the t8 cross section and the 3'-jy cross section, etc. The distance between the goodness of fit line and the reference point is Each RX"rX+ R, +r.

Therefore, from these, 0...
・(6) In this case, unless the 0 portions of the denominators of the first and second terms have the same shape, it cannot be transformed into an explicit function with respect to t. Therefore, mxn (Rx/ax, Ry/ay) = Sm
-...(calculate 71, O<s<s,
Select an appropriate value of S from the range of R,-d, R,-
d.

=S...(8) Find d,, d, which makes a x a y. That is, d, = l(X
−a, s ・=・(9) dν
=Ry-ayS...(lO) is determined, and the function of the isofitness line is transformed as shown in the following equation.

(fR, -d,) /ax+ f2 in t 120
...(1
1) Here, - in decoding is x, y is the positive side, + is x,
y represents the negative side part.

(12) The geometrical meaning of this equation is that, as shown in FIG. 3, it is a shape that is approximated by adding straight line parts to the original ellipse. From this formula...(13) S+17 Hill Seal "〉0...(14) However, when lxl<d, calculate as x=d-, and when lyl<dy, calculate as y=dy. Also, the second As shown,
If the isofitness line of interest is outside the ellipse of fitness l, the values in parentheses are positive.

Similarly, when the isofitness line of interest is inside an ellipse with a goodness of fit of 1, exactly the same equation can be obtained. However, in this case, when any point (x, y) exists in the straight line portion of the approximated isofitness line, lx l <d.

In addition, lyl<d, and the value in parentheses becomes a positive value, and when it is in the tank, the value in parentheses becomes a negative value.

If the inside (or outside) of the ellipse with fitness level l falls into a region with fitness level 1 in a −-like manner, it is determined based on the above conditions and a value is given.

As an example, the three-dimensional membership function for the character figure "8" as shown in FIG. 4 fal is as follows from equation (14).

■A circle with a center point of f6.14) and a radius of 3.5 has a fitness degree I
Let the range of fitness of the three-dimensional elliptic membership function be [
-1,1] (Figure 5)...(15) ■Fitment of a three-dimensional elliptic membership function whose fitness l is a circle whose center point is (6,6) and radius of 4.5 The range of degrees [-
1, 1] (Fig. 6) (16) As shown in Fig. 7 (c), the following function is created by combining the two three-dimensional elliptic membership functions of ■■ above. Let be the three-dimensional membership function for the above character figure "8".

t = max(t+, tz)
= i17) Here, the reason for expanding the range of fitness to [-1, 1] is to give a - (minus) evaluation value to inappropriate input coordinate values. For example, in the application example described below. This is to ensure that when a pitch-black image signal is input, the total sum of degrees of conformity, which will be described later, becomes a small value.

(B) Conversion of Image Input The input image is read by the image reading device 1, and converted into a binary signal including size correction by the signal converter 2.

To explain with the example shown in FIG. 4 (al), in this case,
As shown in Fig. 4fb), the area containing the character "8" is divided into 12 areas in the Convert it to a signal.Here, the blank area is converted to a binary signal of 0, and the black area is converted to a binary signal of 1.As for size correction,
Make sure that the second black area from the top, bottom, left, and right ends is in the second column from the top, bottom, left, and right ends.

At this time, as shown in Figure 7 (bl), 12X 19
n out of 228 areas; binary signal 1 in 35 areas
(black) is given.

(C) Calculating the goodness of fit of the input image For the membership function t set by tAl, the n black areas (dots with a value of 1 among the binary signals (0, 1)) obtained by fBl are The total fitness T and the average fitness per region S+=T+7n are determined.

In the above example, as shown in Figure 7 fbl fc), n
= Total sum of suitability calculated from the coordinate values of 35 black areas Tl = 28, average suitability per area S, =
28/35=0.8.

Similar to the FD+Dl shape compatibility calculation (C), the total compatibility T2 and average compatibility S2 are calculated when a figure with the most ideal shape is created manually.

In the case of the most ideal "8", as shown in FIG.
Average fitness per region 52 = 38/40 = 0.95
becomes.

(E) Calculation of confidence From the results of fcl and (Dl, the ratio TI of the total fitness
The confidence level is calculated by adding /T2 and the average fitness ratio S, /S2 and dividing by 2.

In the example above, the confidence is becomes.

Note that the ratio T, /T, of the total fitness degree exceeds 1 in some cases, but in that case it is unconditionally set to 1.

That is, m1n(T'l/T2.1).

Next, a second example will be explained.

This is an example of identifying the character figure of rA3 as shown in FIG. 8 (at).

+A+Membership Function Setting As shown in FIG. 8 fc), a three-dimensional membership function for the character figure of "A" is formed by combining three three-dimensional elliptic membership functions.

FB+Image Human Power Conversion As shown in Figure 8 fb), a binary signal of 1 (black) is placed in n-37 areas out of 12×19=228 areas.

Calculation of the fitness of the fcl input image + 7 for the membership function set in A), the sum T1 of the fitness of the n black areas obtained in (B), and the average fitness per area S1 = Find Go, /n.

In this example, n
The total suitability of the 37 black areas T, = 30, and the average suitability per area S, = 30/37 = 0.81.

fD) Calculating the degree of conformity of the ideal shape In the case of the most ideal “A”, as shown in Figure 8 fD) (c), the sum of the degrees of conformity of the 40 black areas T 2 =
:38, average fitness per region 52 = 38/400
It becomes 95.

fE) Calculation of confidence From the results of fcl and (Dl, the ratio T of the total fitness degree,
The confidence level is calculated by adding /T2 and the average fitness ratio S+/Sz and dividing the result by 2.

In the example above, the confidence is becomes.

In this procedure, (A) is performed at the initial device design stage, and FDi only needs to be calculated in advance, so when repeating pattern recognition is performed, (Bl, (C1i
The number in which step E) was performed is sufficient.

Ultimately, if the confidence exceeds a certain threshold value, λ is determined to match a specific shape. When the shape to be recognized is compared with a plurality of specific shapes and classified into "C shapes," the shape with the highest degree of certainty is classified.

[Effects of the Invention] As described above, according to the present invention, by using a three-dimensional membership function to express an image force represented by two input variables using a single formula, By setting a fuzzy region with a boundary of an arbitrary shape (for example, an ellipse), highly accurate pattern recognition can be achieved.

To provide a device that can efficiently perform pattern recognition using images held by humans without requiring laborious work such as learning.

Just change the membership function parameters accordingly.
Since the shape of the figure to be recognized can be easily and artificially changed, a versatile and easy-to-use slam recognition device can be obtained.

Since a degree of certainty is given to the recognition result, the range of effective use as a pattern recognition device is expanded.

By using a three-dimensional membership function, it is possible to calculate the confidence level of the recognition result based on the degree of fitness, so it is possible to use decision support systems and fuzzy experts that determine actions using the shape and confidence level obtained as the recognition result as indicators. 1
- Can be used as an input part of the system.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of a fuzzy pattern recognition device according to an embodiment of the present invention. FIG. 2 is a block diagram showing an example of a three-dimensional membership function used in the present invention. FIG. Explanatory diagrams of the case where a straight line part is added to the membership function, Figure 4 (a) (b
) is a diagram showing an example of a character figure to be identified in the embodiment of the present invention and a binary signal corresponding to the shape; FIGS. 5 and 6 are diagrams showing the shape of a three-dimensional membership function used in the embodiment; FIG. 7 is an explanatory diagram of the procedure for identifying the character/figure shown in FIG. 4, and FIG. 8 is an explanatory diagram of the procedure for identifying another character/figure. 1... Image signal input unit 2... Signal conversion unit, 3... Membership function setting unit, 4... Arithmetic processing unit, 5-... Recognition result output unit, 6... Input means.

Claims (2)

[Claims] (1) In a method of performing pattern recognition based on a membership function that defines a fuzzy set, inputting an image signal corresponding to an image to be recognized, converting the image signal into a binary signal corresponding to the image, The system is characterized in that a confidence level as an evaluation value of the image is calculated by comparing the binary signal and a figure with an ambiguous boundary based on a combination of fitness degrees of membership functions of a fuzzy set set in advance. Fuzzy pattern recognition method. (2) A device that performs pattern recognition based on a membership function that defines a fuzzy set, comprising: a signal input unit that inputs an image signal obtained by reading an image to be recognized; A signal conversion unit that converts into a value signal, and a membership function setting that sets one or more three-dimensional membership functions that define a fuzzy set, and forms a figure with ambiguous boundaries by combining the goodness of fit of the membership functions. and an arithmetic unit that calculates and outputs a confidence level as an evaluation value of the image by comparing the binary signal and the figure using a combination of goodness of fit of the membership function as an index. A fuzzy pattern recognition device characterized by: